CONUS404 Spatial Aggregation

CONUS404 Spatial Aggregation#

Using Xarray, GeoPandas and Sparse

Goal: Spatially aggregate a model data variable conservatively, i.e. by exactly partitioning each grid cell into the precise region boundaries.

Approach:

Represent both the original model grid and target grid as GeoPandas GeoSeries objects (with vector geometries)
Compute their area overlay and turn it into a sparse matrix
Perform matrix multiplication on the full Xarray dataset (with a time dimension)

It is quite fast and transparent.

%xmode minimal
import os
import xarray as xr
import geopandas as gp
import pandas as pd
import numpy as np
import sparse

import hvplot.pandas
import hvplot.xarray
import dask
import cf_xarray

from pynhd import NLDI, WaterData
import intake
import cartopy.crs as ccrs
from shapely.geometry import Polygon

Exception reporting mode: Minimal

Open CONUS404 dataset from Intake Catalog#

Select on-prem dataset from /caldera if running on prem (Denali/Tallgrass)
Select cloud/osn data on S3 if running elsewhere.

**NOTE This notebook reads data from S3 (options ending in -cloud), so you will be pulling data from a requester pays bucket. This means you have to set up your AWS credentials, else we won’t be able to load the data. Please note that reading the -cloud data from S3 may incur charges if you are reading data outside of the us-west-2 region or running the notebook outside of the cloud altogether. The following code cell will help set up your AWS credentials:

os.environ['AWS_PROFILE'] = 'default'
%run ../environment_set_up/Help_AWS_Credentials.ipynb

# open the hytest data intake catalog
hytest_cat = intake.open_catalog("https://raw.githubusercontent.com/hytest-org/hytest/main/dataset_catalog/hytest_intake_catalog.yml")
list(hytest_cat)

['conus404-drb-eval-tutorial-catalog',
 'nhm-v1.0-daymet-catalog',
 'nhm-v1.1-c404-bc-catalog',
 'nhm-v1.1-gridmet-catalog',
 'conus404-catalog',
 'nwis-streamflow-usgs-gages-onprem',
 'nwis-streamflow-usgs-gages-cloud',
 'nwm21-streamflow-usgs-gages-onprem',
 'nwm21-streamflow-usgs-gages-cloud',
 'nwm21-streamflow-cloud',
 'nwm21-scores',
 'lcmap-cloud',
 'rechunking-tutorial-cloud']

# open the conus404 sub-catalog
cat = hytest_cat['conus404-catalog']
list(cat)

['conus404-hourly-onprem',
 'conus404-hourly-cloud',
 'conus404-hourly-osn',
 'conus404-daily-diagnostic-onprem',
 'conus404-daily-diagnostic-cloud',
 'conus404-daily-onprem',
 'conus404-daily-cloud',
 'conus404-monthly-onprem',
 'conus404-monthly-cloud']

## Select the dataset you want to read into your notebook and preview its metadata
dataset = 'conus404-daily-cloud' 
cat[dataset]

conus404-daily-cloud:
  args:
    consolidated: true
    storage_options:
      requester_pays: true
    urlpath: s3://nhgf-development/conus404/conus404_daily_202210.zarr/
  description: CONUS404 40 years of daily values for subset of model output variables
    derived from hourly values. This dataset is stored on AWS S3 cloud storage in
    a requester-pays bucket. You can work with this data for free if your workflow
    is running in the us-west-2 region, but you will be charged according to AWS S3
    pricing (https://aws.amazon.com/s3/pricing/) to read the data into a workflow
    running outside of the cloud or in a different AWS cloud region.
  driver: intake_xarray.xzarr.ZarrSource
  metadata:
    catalog_dir: https://raw.githubusercontent.com/hytest-org/hytest/main/dataset_catalog/subcatalogs

Parallelize with Dask#

Some of the steps we will take are aware of parallel clustered compute environments using dask. We’re going to start a cluster now so that future steps can take advantage of this ability.

This is an optional step, but speed ups data loading significantly, especially when accessing data from the cloud.

%run ../environment_set_up/Start_Dask_Cluster_Nebari.ipynb
## If this notebook is not being run on Nebari/ESIP, replace the above 
## path name with a helper appropriate to your compute environment.  Examples:
# %run ../environment_set_up/Start_Dask_Cluster_Denali.ipynb
# %run ../environment_set_up/Start_Dask_Cluster_Tallgrass.ipynb

The 'cluster' object can be used to adjust cluster behavior.  i.e. 'cluster.adapt(minimum=10)'
The 'client' object can be used to directly interact with the cluster.  i.e. 'client.submit(func)' 
The link to view the client dashboard is:
>  https://nebari.esipfed.org/gateway/clusters/dev.1df25e118101470bacf900b89955f6c3/status

ds = cat['conus404-daily-cloud'].to_dask()

Load the feature polygons (e.g. here catchment basins)#

Load with geopandas:

%%time
# USGS gage 01482100 Delaware River at Del Mem Bridge at Wilmington De
gage_id = '01482100'
nldi = NLDI()
del_basins = nldi.get_basins(gage_id)
huc12_basins = WaterData('huc12').bygeom(del_basins.geometry[0])

CPU times: user 362 ms, sys: 55.5 ms, total: 417 ms
Wall time: 5.48 s

regions_df = huc12_basins
region_name = 'name'

Check it:

regions_df.plot(column=region_name, figsize=(10,4))

<Axes: >

../_images/3176082a918577af17604d4be4f2a8f61b24134eff2c0a1bfd008f55e1fa7067.png

All geodataframes should have a coordinate reference system. This is important (and sometimes unfamiliar to users coming from the global climate world).

crs_orig = f'EPSG:{regions_df.crs.to_epsg()}'
crs_orig

'EPSG:4326'

Aggregate CONUS404 to feature polygons#

x = 'x'  # projected x coordinate name
y = 'y'  # projected y coordinate name

ds.crs

<xarray.DataArray 'crs' ()>
[1 values with dtype=int32]
Attributes:
    grid_mapping_name:              lambert_conformal_conic
    latitude_of_projection_origin:  39.100006103515625
    longitude_of_central_meridian:  262.0999984741211
    semi_major_axis:                6370000.0
    semi_minor_axis:                6370000.0
    standard_parallel:              [30.0, 50.0]

crs_info = ds.crs
xx = ds.x.values
yy = ds.y.values
globe = ccrs.Globe(ellipse='sphere', semimajor_axis=6370000, semiminor_axis=6370000)
lcc = ccrs.LambertConformal(globe=globe,
                            central_longitude=crs_info.longitude_of_central_meridian, 
                            central_latitude=crs_info.latitude_of_projection_origin,
                            standard_parallels=crs_info.standard_parallel)

lcc_wkt = lcc.to_wkt()

regions_df = regions_df.to_crs(lcc_wkt)

bbox = tuple(regions_df.total_bounds)
bbox

(1751263.2290125347, 214989.48670458575, 1964829.8118029535, 619548.965931173)

Subset gridded model results to bounds of spatial dataframe to save on memory and computation. More useful when the regions_df is much smaller than the footprint of the gridded model

ds = ds.sel(x=slice(bbox[0],bbox[2]), y=slice(bbox[1],bbox[3]))

Now we extract just the horizontal grid information. The dataset has information about the lat and lon bounds of each cell, which we need to create the CONUS404 grid cell polygons.

var = 'T2'  # 2m Temp
var = 'PREC_ACC_NC' # precip

Now we “stack” the data into a single 1D array. This is the first step towards transitioning to pandas.

grid = grid.cf.add_bounds([x, y])

This function creates geometries for a single pair of bounds. It is not fast, but it is fast enough here. Perhaps could be vectorized using pygeos…

def bounds_to_poly(x_bounds, y_bounds):
    return Polygon([
        (x_bounds[0], y_bounds[0]),
        (x_bounds[0], y_bounds[1]),
        (x_bounds[1], y_bounds[1]),
        (x_bounds[1], y_bounds[0])
    ])

We apply this function to each grid cell.

%%time
boxes = xr.apply_ufunc(
    bounds_to_poly,
    points.x_bounds,
    points.y_bounds,
    input_core_dims=[("bounds",),  ("bounds",)],
    output_dtypes=[np.dtype('O')],
    vectorize=True
)

CPU times: user 86.8 ms, sys: 40.5 ms, total: 127 ms
Wall time: 82.6 ms

Finally, we convert to a GeoDataframe, specifying the projected CRS

grid_df= gp.GeoDataFrame(
    data={"geometry": boxes.values, "y": boxes[y], "x": boxes[x]},
    index=boxes.indexes["point"],
    crs=lcc_wkt
)

We will now transform to an area preserving projection. This is imporant because we want to do area-weighted regridding. Here we use the NSIDC EASE-Grid 2.0 grid for the Northern Hemisphere.

crs_area = "EPSG:6931"

regions_df = regions_df.to_crs(crs_area)
grid_df = grid_df.to_crs(crs_area)

grid_df.crs

<Projected CRS: EPSG:6931>
Name: WGS 84 / NSIDC EASE-Grid 2.0 North
Axis Info [cartesian]:
- X[south]: Easting (metre)
- Y[south]: Northing (metre)
Area of Use:
- name: Northern hemisphere.
- bounds: (-180.0, 0.0, 180.0, 90.0)
Coordinate Operation:
- name: US NSIDC EASE-Grid 2.0 North
- method: Lambert Azimuthal Equal Area
Datum: World Geodetic System 1984 ensemble
- Ellipsoid: WGS 84
- Prime Meridian: Greenwich

Key Step: Overlay the two geometries#

This is the magic of geopandas; it can calculate the overlap between the original grid and the regions. It is expensive because it has to compare 64800 grid boxes with 242 regions.

In this dataframe, the latitude and longitude values are from the grid, while all the other columns are from the regions.

overlay = grid_df.overlay(regions_df, keep_geom_type=True)

This is essentially already a sparse matrix mapping one grid space to the other. How sparse?

sparsity = len(overlay) / (len(grid_df) * len(regions_df))
sparsity

0.0024750477952660914

Let’s explore these overlays a little bit

We can verify that each basin’s area is preserved in the overlay operation.

overlay.geometry.area.groupby(overlay[region_name]).sum().nlargest(10)/1e6  # km2

name
Delaware Bay-Deep                         1073.773336
Ashokan Reservoir-Esopus Creek             205.798540
Lower Little Swatara Creek                 194.683255
Lake Wallenpaupack-Wallenpaupack Creek     180.104323
Bear Creek                                 176.743216
Bear Kill                                  174.877729
Upper Mahanoy Creek                        170.342913
Headwaters Paulins Kill River              167.072157
Black Creek                                160.267392
Snitz Creek-Quittapahilla Creek            159.798726
dtype: float64

overlay[overlay[region_name] == "Delaware Bay-Deep"].geometry.plot(edgecolor='k', aspect='equal')

<Axes: >

../_images/2a582d441b4c1538c7e150c7332b8a391d247742e8a93dcab1aaf61d6d5330f9.png

regions_df.geometry.area.groupby(regions_df[region_name]).sum().nlargest(10)

name
Delaware Bay-Deep                         1.073773e+09
Ashokan Reservoir-Esopus Creek            2.057985e+08
Lower Little Swatara Creek                1.946833e+08
Lake Wallenpaupack-Wallenpaupack Creek    1.801043e+08
Bear Creek                                1.767432e+08
Bear Kill                                 1.748777e+08
Upper Mahanoy Creek                       1.703429e+08
Headwaters Paulins Kill River             1.670722e+08
Black Creek                               1.602674e+08
Snitz Creek-Quittapahilla Creek           1.597987e+08
dtype: float64

Calculate the area fraction for each region#

This is another key step. This transform tells us how much of a country’s total area comes from each of the grid cells. This is accurate because we used an area-preserving CRS.

grid_cell_fraction = overlay.geometry.area.groupby(overlay[region_name]).transform(lambda x: x / x.sum())
grid_cell_fraction

     0.006051
     0.007286
     0.000754
     0.004720
     0.014627
          ...   
  0.001910
  0.179560
  0.180305
  0.093585
  0.000193
Length: 6196, dtype: float64

We can verify that these all sum up to one.

grid_cell_fraction.groupby(overlay[region_name]).sum()

name
Alexauken Creek-Delaware River                       1.0
Allegheny Creek-Delaware River                       1.0
Angelica Creek-Schuylkill River                      1.0
Antietam Creek                                       1.0
Appenzell Creek                                      1.0
                                                    ... 
Wrangle Brook                                        1.0
Wright Brook-Headwaters West Brach Delaware River    1.0
Wright Creek-Lehigh River                            1.0
Wyomissing Creek                                     1.0
Yaleville Brook-Susquehanna River                    1.0
Length: 451, dtype: float64

Turn this into a sparse Xarray DataArray#

The first step is making a MultiIndex

multi_index = overlay.set_index([y, x, region_name]).index
df_weights = pd.DataFrame({"weights": grid_cell_fraction.values}, index=multi_index)
df_weights

			weights
y	x	name
216039.100363	1.919902e+06	Delaware Bay-Deep	0.006051
216039.100363	1.923902e+06	Delaware Bay-Deep	0.007286
220039.100363	1.911902e+06	Delaware Bay-Deep	0.000754
	1.915902e+06	Delaware Bay-Deep	0.004720
	1.919902e+06	Delaware Bay-Deep	0.014627
...	...	...	...
616039.100363	1.859902e+06	Mine Kill	0.001910
	1.863902e+06	Mine Kill	0.179560
	1.867902e+06	Mine Kill	0.180305
	1.871902e+06	Mine Kill	0.093585
	1.875902e+06	Mine Kill	0.000193

6196 rows × 1 columns

We can bring this directly into xarray as a 1D Dataset.

ds_weights = xr.Dataset(df_weights)

Now we unstack it into a sparse array.

weights_sparse = ds_weights.unstack(sparse=True, fill_value=0.).weights

Here we can clearly see that this is a sparse matrix, mapping the input space (lat, lon) to the output space (SOVEREIGNT).

Perform Matrix Multiplication#

To regrid the data, we just have to multiply the original precip dataset by this matrix.

#regridded = xr.dot(ds[var], weights_sparse)

#regridded = sparse.einsum('ij,jk->ik', ds[var].data, weights_sparse.data)

Unfortunately, that doesn’t work out of the box, because sparse doesn’t implement einsum (see https://github.com/pydata/sparse/issues/31).

# regridded[0].compute()  # fails

Sparse does implement matmul, so we can use that. But we have to do some reshaping to make it work with our data.

def apply_weights_matmul_sparse(weights, data):

    assert isinstance(weights, sparse.SparseArray)
    assert isinstance(data, np.ndarray)
    data = sparse.COO.from_numpy(data)
    data_shape = data.shape
    # k = nlat * nlon
    n, k = data_shape[0], data_shape[1] * data_shape[2]
    data = data.reshape((n, k))
    weights_shape = weights.shape
    k_, m = weights_shape[0] * weights_shape[1], weights_shape[2]
    assert k == k_
    weights_data = weights.reshape((k, m))

    regridded = sparse.matmul(data, weights_data)
    assert regridded.shape == (n, m)
    return regridded.todense()

Explore timeseries data by region:#

Plot monthly-average time series for two selected HUC12s.

ds_var = var_regridded.sel(name=['Delaware Bay-Deep', 'Black Creek']).resample(time="MS").mean().to_dataset(region_name)

ds_var.hvplot(x='time', grid=True, frame_width=1000)

Explore the mean var by region#

Calculate the average value over all time steps for every HU12

df_mean = var_regridded.to_pandas().mean()
df_mean.name = var
df_mean = pd.DataFrame(df_mean).reset_index()

merged = pd.merge(regions_df, df_mean)

Convert back to geographic coordinates for plotting

crs_geo = 'EPSG:4326'

merged_geo = merged.to_crs(crs_geo)

Holoviz:

merged_geo.hvplot(c=var, geo=True, cmap='viridis_r', frame_width=600, tiles='StamenTerrain', 
               title='CONUS404', alpha=0.7)

Be a good citizen and shut down the cluster#

… Even though it would shut down in 15 minutes with no activity

client.close(); cluster.shutdown()

/home/conda/users/ef214d1986de9477d5223419ba4838d76e9839b19e4fb02c9e208c97cab5a61b-20230329-131713-443303-165-pangeofu/lib/python3.10/site-packages/dask_gateway/client.py:1014: RuntimeWarning: coroutine 'rpc.close_rpc' was never awaited
  self.scheduler_comm.close_rpc()